mirror of
https://github.com/TheTom/turboquant_plus.git
synced 2026-10-22 02:53:13 -05:00
693a391333
Codex fixes: - wall_ms → wall_sec (BSD date doesn't support %3N milliseconds) - Linux physical cores: use 'core id' count instead of processor count - Remove unused $JSONFILE variable (caused crash on unset -u) - Fix Metal shader comment accuracy (signs can mix per element) Gemini/doc review fixes: - Add inflection point detection (find_decode_inflection) - Add 1K context measurement flagging (flag_unreliable_measurements) - Add has_tensor impact to prediction model (+30% penalty) - Use exponential cache fraction model (matches real data better) - 4 new tests: inflection detection, unreliable measurement flagging 255 total tests passing (144 core + 71 diag + 40 replay). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-Authored-By: tturney@psyguard.ai